Openai says she ignored expert expert fears when she put up an update to the pioneering artificial intelligence model Chatgpt, which made her excessively acceptable.
The company released an update of the GPT -4O model on April 25 that made it “more clearly clear”, which fell after three days due to safety concerns, Openaii He said In the May 2 blog publication.
Chatgpt maker said its new models are subject to safety and behavior, and its internal experts “spend” a great time to interact with every new model before launch, “and they go to the issues that other tests lack.
During the process of reviewing the last model before it was publicly prayed, Openai said that “some experts test have indicated that the behavior of the model is” a little “” poetry “, but they decided to launch” because of the positive signals of the users who tried the model. “
“Unfortunately, this was the wrong call,” the company admitted. “The qualitative assessments glimpsed something important, and we should pay close attention. They were picking up in a blind place in Evals and our other standards.”
On a large scale, the text -based artificial intelligence models are trained by giving them accurate responses or significantly classified by their coaches. Some rewards are given heavier weighting, which affects how the model responds.
Obayye said that the introduction of the user notes bonus signal is “the basic reward signal for the model, which kept the verification”, which led to being more binding.
He added: “The user’s notes in particular may sometimes prefer more acceptable responses, and it is possible that the transformation we have seen.”
Openai is now verifying
After publishing the updated artificial intelligence model, Chatgpt users via the Internet complained of his tendency to shower on any idea presented, regardless of its misfortune, which led waiver In a blog post on April 29, “It was great or acceptable.”
For example, one of the users told Chatgpt that he wanted to start a business that sells ice online, which guarantees the sale of regular old water to customers for freezing.
In the last post -death, he said that this behavior of artificial intelligence may pose a danger, especially with issues such as mental health.
Openai said: “People have begun to use Chatgpt for profound personal advice – something we haven’t seen even a year ago,” Openai said. “With the development of artificial intelligence and society, it became clear that we need to deal with this state of use very carefully.”
Related to: Cold encryption users with artificial intelligence with their wallets
The company said it had discussed the risk of SYCOPHANCY “for a period of time”, but it was not explicitly marked for the internal test, and it had no specific methods to track Sycophance.
Now, he will look forward to adding “SYCOPHANCY” assessments by controlling its safety review process to “formally look at behavior problems” and prohibit the launch of a model if it provides problems.
Openai also admitted that he had not announced the latest model because he expected to be a “somewhat hidden update”, which he pledged to change.
“There is nothing like a” small “launch.” “We will try to communicate the accurate changes that can change useful how people interact with Chatgpt,” the company wrote.
Artificial Intelligence Eye: Crypto AI symbols rise by 34 %, why did you see such a kiss
Related Articles
